From eb57123e0d98a2095d499b4ba7c9b09d30945ec2 Mon Sep 17 00:00:00 2001
From: TiM <tim@nuimarkets.com>
Date: Tue, 1 Sep 2026 08:34:49 +1200
Subject: [PATCH] Mirror `readinto` from the pyarrow stream onto `ArrowFile`

`ArrowFile` mirrors a fixed method list from the pyarrow stream it wraps and
`readinto` is not on it, though `pyarrow.NativeFile` implements it. That leaves
it the one file class here without one, so code written against the
`AbstractBufferedFile` contract breaks on an Arrow-backed filesystem and
nowhere else.

`LibArchiveFileSystem` is the in-repo case, and it fails silently. Its
`custom_reader` calls `readinto` from a ctypes callback, which swallows the
`AttributeError`, so libarchive sees a zero-length read and `ls()` returns an
empty archive instead of raising. Buffered reads at least raise: a sized `read`,
`peek` and `read1` each fill the buffer through `readinto`, while an unsized
`read()` falls back to `read(-1)` and is unaffected.

Gzip is how it turned up. `fsspec.compression` registers isal's `IGzipFile` as
the `gzip` codec when `isal` imports, and that reader decompresses through
`readinto`, so reading a `.gz` file through an Arrow-backed filesystem fails in
an environment that happens to carry the package and reads fine everywhere
else.
---
 fsspec/implementations/arrow.py            |  1 +
 fsspec/implementations/tests/test_arrow.py | 36 ++++++++++++++++++++++
 2 files changed, 37 insertions(+)

diff --git a/fsspec/implementations/arrow.py b/fsspec/implementations/arrow.py
index c312cd4a5..d17351441 100644
--- a/fsspec/implementations/arrow.py
+++ b/fsspec/implementations/arrow.py
@@ -227,6 +227,7 @@ def get_file(self, rpath, lpath, **kwargs):
     "stream",
     [
         "read",
+        "readinto",
         "seek",
         "tell",
         "write",
diff --git a/fsspec/implementations/tests/test_arrow.py b/fsspec/implementations/tests/test_arrow.py
index 52033c938..acbfea379 100644
--- a/fsspec/implementations/tests/test_arrow.py
+++ b/fsspec/implementations/tests/test_arrow.py
@@ -1,3 +1,4 @@
+import io
 import secrets
 
 import pytest
@@ -271,6 +272,41 @@ def test_seekable(fs, remote_dir):
             file.seek(5)
 
 
+def test_readinto(fs, remote_dir):
+    data = b"dvc.org"
+
+    with fs.open(remote_dir + "/a.txt", "wb") as stream:
+        stream.write(data)
+
+    for seekable in [True, False]:
+        with fs.open(remote_dir + "/a.txt", "rb", seekable=seekable) as file:
+            buffer = bytearray(3)
+            assert file.readinto(buffer) == 3
+            assert bytes(buffer) == data[:3]
+
+
+@pytest.mark.parametrize(
+    "read",
+    [
+        lambda buffered: buffered.read(3),
+        lambda buffered: buffered.peek(3)[:3],
+        lambda buffered: buffered.read1(3),
+    ],
+    ids=["read", "peek", "read1"],
+)
+def test_readinto_supports_a_buffered_reader(fs, remote_dir, read):
+    # Each of these fills a buffer io.BufferedReader owns, so they go through
+    # readinto rather than read(); an unsized read() does not. A gzip reader takes
+    # the peek()/read1() path.
+    data = b"dvc.org"
+
+    with fs.open(remote_dir + "/a.txt", "wb") as stream:
+        stream.write(data)
+
+    with fs.open(remote_dir + "/a.txt", "rb") as file:
+        assert read(io.BufferedReader(file)) == data[:3]
+
+
 def test_get_kwargs_from_urls_hadoop_fs():
     kwargs = HadoopFileSystem._get_kwargs_from_urls(
         "hdfs://user@localhost:8020/?replication=2"